top of page
Writer's pictureRémi Coscoy

Discover Kiadam's best practices for use in your deep learning computer vision project

Updated: Aug 29, 2023

Using a new tool can be overwhelming at first. If you have any doubt about what setting to use or which images to upload, this table contains all the answers you need.

Some settings need a bit more explanations, click on the link to read a more detailed paragraph.

Setting

Recommendation

Commentary

Image Size

640x640

Ultralytics' Yolo default Image Size

Number of generated Images

1000

Size of objects relative to the final image

Specific to your Usecase

Number of objects in the final Image

Specific to your Usecase

Percent Only Background

Between 1% and 10%

Test/Train/Val split

80% Train 20% Val

We recommend a real life testing set when possible

Labeled Objects

Between 10 and 15 per object

Get as many angles of your object as possible

Background Images

Specific to your Usecase

Additional Backgrounds

Only if you don't have any other backgrounds

Image backgrounds will usually perform better

Transformations

Specific to your Usecase

How to choose the size of your objects


Try to reproduce the variations you will encounter in your production environments

Will your object always be the main focus of the image? Select a size range from 70% to 90% combined with a single object per image

If you object will be an element amongst other in the image, you should select a size range from 10% to 30%.

If you might encounter all of the above or are unsure, select a wide size range, such as from 10% to 90%

See a more detailed explanation here


How to choose the number of objects in the final image


Ask yourself how many objects will there be per image in your production environment.

If the answer is for sure one, put only one object per image.

If there might be several objects per Image, set that more objects should be in each Image.

Beware however of the size range : Having 5 images per object, each taking up 90% of total image size will result in a completely covered background and a high overlap between objects.

We recommend lowering the size range when several objects might be in the image.


How to choose your Background Images


Do you know for sure the background or backgrounds you will encounter in your production environments? Set various Images from those places as background Images

If you want to generalize your training set, we recommend using a collection of varied background such as this one

See a more detailed explanation here


How to choose your Transformations


This is by far the most difficult step and where the quality of your training set will vary the most. There is no magic formula, and the correct combinations will depend on your use-case.

Try to reproduce as closely as possible the variations you will encounter in production.

Skew, Rotate and Flip are recommended as long as it makes sense for your object. (Good for a can of soda, but having upside down animals in your training dataset might deteriorate your precision as you will most likely encounter animals in the "correct" orientation in production)

Overall, be parsimonious with the transformations : too many and your object might become unrecognizable.

We recommend trying several combinations and see which one sticks.


41 views
bottom of page